Finally got the fragment program fully working for bicubic filtering on r5xx.

1 files changed, 116 insertions, 70 deletions

diff --git a/src/radeon_textured_videofuncs.c b/src/radeon_textured_videofuncs.c
index 58bb0ab4..5b7b7875 100644
--- a/src/radeon_textured_videofuncs.c
+++ b/src/radeon_textured_videofuncs.c
@@ -414,7 +414,7 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 		OUT_VIDEO_REG(R300_RS_INST_COUNT, R300_INST_COUNT_RS(1) | R300_TX_OFFSET_RS(6));
 
 		/* Pixel stack frame size. */
-		OUT_VIDEO_REG(R500_US_PIXSIZE, R500_PIX_SIZE(15));
+		OUT_VIDEO_REG(R500_US_PIXSIZE, R500_PIX_SIZE(8));
 
 		/* FP length. */
 		OUT_VIDEO_REG(R500_US_CODE_ADDR, (R500_US_CODE_START_ADDR(0) |
@@ -507,7 +507,9 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGB_B_SWIZ_B_G));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(4) |
 						   R500_ALPHA_OP_MAD |
+						   R500_ALPHA_SEL_A_SRC0 |
 						   R500_ALPHA_SWIZ_A_A |
+						   R500_ALPHA_SEL_B_SRC1 |
 						   R500_ALPHA_SWIZ_B_G));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(4) |
 						   R500_ALU_RGBA_OP_MAD |
@@ -539,7 +541,9 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGB_MOD_B_NEG));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(5) |
 						   R500_ALPHA_OP_MAD |
+						   R500_ALPHA_SEL_A_SRC0 |
 						   R500_ALPHA_SWIZ_A_A |
+						   R500_ALPHA_SEL_B_SRC1 |
 						   R500_ALPHA_SWIZ_B_R |
 						   R500_ALPHA_MOD_B_NEG));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(5) |
@@ -573,7 +577,9 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGB_B_SWIZ_B_G));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(6) |
 						   R500_ALPHA_OP_MAD |
+						   R500_ALPHA_SEL_A_SRC0 |
 						   R500_ALPHA_SWIZ_A_A |
+						   R500_ALPHA_SEL_B_SRC1 |
 						   R500_ALPHA_SWIZ_B_G));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(6) |
 						   R500_ALU_RGBA_OP_MAD |
@@ -607,7 +613,9 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGB_B_SWIZ_B_G));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(7) |
 						   R500_ALPHA_OP_MAD |
+						   R500_ALPHA_SEL_A_SRC0 |
 						   R500_ALPHA_SWIZ_A_A |
+						   R500_ALPHA_SEL_B_SRC1 |
 						   R500_ALPHA_SWIZ_B_G));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(7) |
 						   R500_ALU_RGBA_OP_MAD |
@@ -642,8 +650,11 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGB_MOD_B_NEG));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(4) |
 						   R500_ALPHA_OP_MAD |
+						   R500_ALPHA_SEL_A_SRC0 |
 						   R500_ALPHA_SWIZ_A_A |
-						   R500_ALPHA_SWIZ_B_R));
+						   R500_ALPHA_SEL_B_SRC1 |
+						   R500_ALPHA_SWIZ_B_R |
+						   R500_ALPHA_MOD_B_NEG));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(4) |
 						   R500_ALU_RGBA_OP_MAD |
 						   R500_ALU_RGBA_SEL_C_SRC2 |
@@ -677,8 +688,11 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGB_MOD_B_NEG));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(5) |
 						   R500_ALPHA_OP_MAD |
+						   R500_ALPHA_SEL_A_SRC0 |
 						   R500_ALPHA_SWIZ_A_A |
-						   R500_ALPHA_SWIZ_B_R));
+						   R500_ALPHA_SEL_B_SRC1 |
+						   R500_ALPHA_SWIZ_B_R |
+						   R500_ALPHA_MOD_B_NEG));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(5) |
 						   R500_ALU_RGBA_OP_MAD |
 						   R500_ALU_RGBA_SEL_C_SRC2 |
@@ -708,7 +722,9 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(4) |
 						   R500_ALPHA_OP_MAD |
 						   R500_ALPHA_SWIZ_A_1 |
-						   R500_ALPHA_SWIZ_B_A));
+						   R500_ALPHA_SEL_B_SRC1 |
+						   R500_ALPHA_SWIZ_B_A |
+						   R500_ALPHA_MOD_B_NEG));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(4) |
 						   R500_ALU_RGBA_OP_MAD |
 						   R500_ALU_RGBA_SEL_C_SRC2 |
@@ -738,7 +754,9 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(5) |
 						   R500_ALPHA_OP_MAD |
 						   R500_ALPHA_SWIZ_A_1 |
-						   R500_ALPHA_SWIZ_B_A));
+						   R500_ALPHA_SEL_B_SRC1 |
+						   R500_ALPHA_SWIZ_B_A |
+						   R500_ALPHA_MOD_B_NEG));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(5) |
 						   R500_ALU_RGBA_OP_MAD |
 						   R500_ALU_RGBA_SEL_C_SRC2 |
@@ -768,7 +786,9 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(6) |
 						   R500_ALPHA_OP_MAD |
 						   R500_ALPHA_SWIZ_A_1 |
-						   R500_ALPHA_SWIZ_B_A));
+						   R500_ALPHA_SEL_B_SRC1 |
+						   R500_ALPHA_SWIZ_B_A |
+						   R500_ALPHA_MOD_B_NEG));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(6) |
 						   R500_ALU_RGBA_OP_MAD |
 						   R500_ALU_RGBA_SEL_C_SRC2 |
@@ -798,7 +818,9 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(7) |
 						   R500_ALPHA_OP_MAD |
 						   R500_ALPHA_SWIZ_A_1 |
-						   R500_ALPHA_SWIZ_B_A));
+						   R500_ALPHA_SEL_B_SRC1 |
+						   R500_ALPHA_SWIZ_B_A |
+						   R500_ALPHA_MOD_B_NEG));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(7) |
 						   R500_ALU_RGBA_OP_MAD |
 						   R500_ALU_RGBA_SEL_C_SRC2 |
@@ -908,37 +930,41 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, 0x00000000);
 
 		/* LRP temp3, temp1.zzzz, temp3, temp5 ->
-		 * - ADD temp6, temp3, -temp5
-		 * - MAD temp3, temp1.zzzz, temp6, temp5 */
+		 * - PRESUB temps, 1 - temp1.zzzz
+		 * - MUL temp5, temps, temp5
+		 * - MAD temp3, temp1.zzzz, temp3, temp5 */
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_ALU |
 						   R500_INST_TEX_SEM_WAIT |
 						   R500_INST_RGB_WMASK_R |
 						   R500_INST_RGB_WMASK_G |
 						   R500_INST_RGB_WMASK_B |
 						   R500_INST_ALPHA_WMASK));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR1(5) |
-						   R500_RGB_ADDR2(7)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR1(5) |
-						   R500_ALPHA_ADDR2(7)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_R_SWIZ_A_1 |
-						   R500_ALU_RGB_G_SWIZ_A_1 |
-						   R500_ALU_RGB_B_SWIZ_A_1 |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(3) |
+						   R500_RGB_SRCP_OP_1_MINUS_RGB0 |
+						   R500_RGB_ADDR1(7)));
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(3) |
+						   R500_ALPHA_SRCP_OP_1_MINUS_A0 |
+						   R500_ALPHA_ADDR1(7)));
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_SEL_A_SRCP |
+						   R500_ALU_RGB_R_SWIZ_A_B |
+						   R500_ALU_RGB_G_SWIZ_A_B |
+						   R500_ALU_RGB_B_SWIZ_A_B |
 						   R500_ALU_RGB_SEL_B_SRC1 |
 						   R500_ALU_RGB_R_SWIZ_B_R |
 						   R500_ALU_RGB_G_SWIZ_B_G |
 						   R500_ALU_RGB_B_SWIZ_B_B));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(8) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(7) |
 						   R500_ALPHA_OP_MAD |
-						   R500_ALPHA_SWIZ_A_1 |
+						   R500_ALPHA_SEL_A_SRCP |
+						   R500_ALPHA_SWIZ_A_B |
+						   R500_ALPHA_SEL_B_SRC1 |
 						   R500_ALPHA_SWIZ_B_A));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(8) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(7) |
 						   R500_ALU_RGBA_OP_MAD |
-						   R500_ALU_RGBA_SEL_C_SRC2 |
-						   R500_ALU_RGBA_R_SWIZ_R |
-						   R500_ALU_RGBA_G_SWIZ_G |
-						   R500_ALU_RGBA_B_SWIZ_B |
-						   R500_ALU_RGBA_A_SWIZ_A |
-						   R500_ALU_RGBA_ALPHA_MOD_C_NEG));
+						   R500_ALU_RGBA_R_SWIZ_0 |
+						   R500_ALU_RGBA_G_SWIZ_0 |
+						   R500_ALU_RGBA_B_SWIZ_0 |
+						   R500_ALU_RGBA_A_SWIZ_0));
 		/* 2nd inst */
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_ALU |
 						   R500_INST_RGB_WMASK_R |
@@ -946,10 +972,10 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_INST_RGB_WMASK_B |
 						   R500_INST_ALPHA_WMASK));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(3) |
-						   R500_RGB_ADDR1(8) |
+						   R500_RGB_ADDR1(5) |
 						   R500_RGB_ADDR2(7)));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(3) |
-						   R500_ALPHA_ADDR1(8) |
+						   R500_ALPHA_ADDR1(5) |
 						   R500_ALPHA_ADDR2(7)));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_SEL_A_SRC0 |
 						   R500_ALU_RGB_R_SWIZ_A_B |
@@ -961,7 +987,9 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGB_B_SWIZ_B_B));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(5) |
 						   R500_ALPHA_OP_MAD |
+						   R500_ALPHA_SEL_A_SRC0 |
 						   R500_ALPHA_SWIZ_A_B |
+						   R500_ALPHA_SEL_B_SRC1 |
 						   R500_ALPHA_SWIZ_B_A));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(5) |
 						   R500_ALU_RGBA_OP_MAD |
@@ -972,36 +1000,40 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGBA_A_SWIZ_A));
 
 		/* LRP temp2, temp1.zzzz, temp2, temp4 ->
-		 * - ADD temp6, temp2, -temp4
-		 * - MAD temp2, temp1.zzzz, temp6, temp4 */
+		 * - PRESUB temps, 1 - temp1.zzzz
+		 * - ADD temp4, temps, temp4
+		 * - MAD temp2, temp1.zzzz, temp2, temp4 */
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_ALU |
 						   R500_INST_RGB_WMASK_R |
 						   R500_INST_RGB_WMASK_G |
 						   R500_INST_RGB_WMASK_B |
 						   R500_INST_ALPHA_WMASK));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR1(4) |
-						   R500_RGB_ADDR2(6)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR1(4) |
-						   R500_ALPHA_ADDR2(6)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_R_SWIZ_A_1 |
-						   R500_ALU_RGB_G_SWIZ_A_1 |
-						   R500_ALU_RGB_B_SWIZ_A_1 |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(3) |
+						   R500_RGB_SRCP_OP_1_MINUS_RGB0 |
+						   R500_RGB_ADDR1(6)));
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(3) |
+						   R500_ALPHA_SRCP_OP_1_MINUS_A0 |
+						   R500_ALPHA_ADDR1(6)));
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_SEL_A_SRCP |
+						   R500_ALU_RGB_R_SWIZ_A_B |
+						   R500_ALU_RGB_G_SWIZ_A_B |
+						   R500_ALU_RGB_B_SWIZ_A_B |
 						   R500_ALU_RGB_SEL_B_SRC1 |
 						   R500_ALU_RGB_R_SWIZ_B_R |
 						   R500_ALU_RGB_G_SWIZ_B_G |
 						   R500_ALU_RGB_B_SWIZ_B_B));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(8) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(6) |
 						   R500_ALPHA_OP_MAD |
-						   R500_ALPHA_SWIZ_A_1 |
+						   R500_ALPHA_SEL_A_SRCP |
+						   R500_ALPHA_SWIZ_A_B |
+						   R500_ALPHA_SEL_B_SRC1 |
 						   R500_ALPHA_SWIZ_B_A));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(8) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(6) |
 						   R500_ALU_RGBA_OP_MAD |
-						   R500_ALU_RGBA_SEL_C_SRC2 |
-						   R500_ALU_RGBA_R_SWIZ_R |
-						   R500_ALU_RGBA_G_SWIZ_G |
-						   R500_ALU_RGBA_B_SWIZ_B |
-						   R500_ALU_RGBA_A_SWIZ_A |
-						   R500_ALU_RGBA_ALPHA_MOD_C_NEG));
+						   R500_ALU_RGBA_R_SWIZ_0 |
+						   R500_ALU_RGBA_G_SWIZ_0 |
+						   R500_ALU_RGBA_B_SWIZ_0 |
+						   R500_ALU_RGBA_A_SWIZ_0));
 		/* 2nd inst */
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_ALU |
 						   R500_INST_RGB_WMASK_R |
@@ -1009,10 +1041,10 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_INST_RGB_WMASK_B |
 						   R500_INST_ALPHA_WMASK));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(3) |
-						   R500_RGB_ADDR1(8) |
+						   R500_RGB_ADDR1(4) |
 						   R500_RGB_ADDR2(6)));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(3) |
-						   R500_ALPHA_ADDR1(8) |
+						   R500_ALPHA_ADDR1(4) |
 						   R500_ALPHA_ADDR2(6)));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_SEL_A_SRC0 |
 						   R500_ALU_RGB_R_SWIZ_A_B |
@@ -1024,7 +1056,9 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGB_B_SWIZ_B_B));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(4) |
 						   R500_ALPHA_OP_MAD |
+						   R500_ALPHA_SEL_A_SRC0 |
 						   R500_ALPHA_SWIZ_A_B |
+						   R500_ALPHA_SEL_B_SRC1 |
 						   R500_ALPHA_SWIZ_B_A));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(4) |
 						   R500_ALU_RGBA_OP_MAD |
@@ -1035,36 +1069,40 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGBA_A_SWIZ_A));
 
 		/* LRP output, temp0.zzzz, temp3, temp2 ->
-		 * - ADD temp6, temp3, -temp2
-		 * - MAD output, temp0.zzzz, temp6, temp2 */
+		 * - PRESUB temps, 1 - temp0.zzzz
+		 * - MUL temp2, temps, temp2
+		 * - MAD output, temp0.zzzz, temp3, temp2 */
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_ALU |
 						   R500_INST_RGB_WMASK_R |
 						   R500_INST_RGB_WMASK_G |
 						   R500_INST_RGB_WMASK_B |
 						   R500_INST_ALPHA_WMASK));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR1(5) |
-						   R500_RGB_ADDR2(4)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR1(5) |
-						   R500_ALPHA_ADDR2(4)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_R_SWIZ_A_1 |
-						   R500_ALU_RGB_G_SWIZ_A_1 |
-						   R500_ALU_RGB_B_SWIZ_A_1 |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(2) |
+						   R500_RGB_SRCP_OP_1_MINUS_RGB0 |
+						   R500_RGB_ADDR1(4)));
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(2) |
+						   R500_ALPHA_SRCP_OP_1_MINUS_A0 |
+						   R500_ALPHA_ADDR1(4)));
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_SEL_A_SRCP |
+						   R500_ALU_RGB_R_SWIZ_A_B |
+						   R500_ALU_RGB_G_SWIZ_A_B |
+						   R500_ALU_RGB_B_SWIZ_A_B |
 						   R500_ALU_RGB_SEL_B_SRC1 |
 						   R500_ALU_RGB_R_SWIZ_B_R |
 						   R500_ALU_RGB_G_SWIZ_B_G |
 						   R500_ALU_RGB_B_SWIZ_B_B));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(8) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(4) |
 						   R500_ALPHA_OP_MAD |
-						   R500_ALPHA_SWIZ_A_1 |
+						   R500_ALPHA_SEL_A_SRCP |
+						   R500_ALPHA_SWIZ_A_B |
+						   R500_ALPHA_SEL_B_SRC1 |
 						   R500_ALPHA_SWIZ_B_A));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(8) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(4) |
 						   R500_ALU_RGBA_OP_MAD |
-						   R500_ALU_RGBA_SEL_C_SRC2 |
-						   R500_ALU_RGBA_R_SWIZ_R |
-						   R500_ALU_RGBA_G_SWIZ_G |
-						   R500_ALU_RGBA_B_SWIZ_B |
-						   R500_ALU_RGBA_A_SWIZ_A |
-						   R500_ALU_RGBA_ALPHA_MOD_C_NEG));
+						   R500_ALU_RGBA_R_SWIZ_0 |
+						   R500_ALU_RGBA_G_SWIZ_0 |
+						   R500_ALU_RGBA_B_SWIZ_0 |
+						   R500_ALU_RGBA_A_SWIZ_0));
 		/* 2nd inst */
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_OUT |
 						   R500_INST_LAST |
@@ -1072,12 +1110,16 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_INST_RGB_WMASK_R |
 						   R500_INST_RGB_WMASK_G |
 						   R500_INST_RGB_WMASK_B |
-						   R500_INST_ALPHA_WMASK));
+						   R500_INST_ALPHA_WMASK |
+						   R500_INST_RGB_OMASK_R |
+						   R500_INST_RGB_OMASK_G |
+						   R500_INST_RGB_OMASK_B |
+						   R500_INST_ALPHA_OMASK));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(2) |
-						   R500_RGB_ADDR1(8) |
+						   R500_RGB_ADDR1(5) |
 						   R500_RGB_ADDR2(4)));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(2) |
-						   R500_ALPHA_ADDR1(8) |
+						   R500_ALPHA_ADDR1(5) |
 						   R500_ALPHA_ADDR2(4)));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_SEL_A_SRC0 |
 						   R500_ALU_RGB_R_SWIZ_A_B |
@@ -1087,10 +1129,14 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGB_R_SWIZ_B_R |
 						   R500_ALU_RGB_G_SWIZ_B_G |
 						   R500_ALU_RGB_B_SWIZ_B_B));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_OP_MAD |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(0) |
+						   R500_ALPHA_OP_MAD |
+						   R500_ALPHA_SEL_A_SRC0 |
 						   R500_ALPHA_SWIZ_A_B |
+						   R500_ALPHA_SEL_B_SRC1 |
 						   R500_ALPHA_SWIZ_B_A));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_OP_MAD |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(0) |
+						   R500_ALU_RGBA_OP_MAD |
 						   R500_ALU_RGBA_SEL_C_SRC2 |
 						   R500_ALU_RGBA_R_SWIZ_R |
 						   R500_ALU_RGBA_G_SWIZ_G |